A remote sensing image building change detection method based on interactive attention enhancement

Abstract

The application discloses a remote sensing image building change detection method based on interactive attention enhancement, belongs to the field of remote sensing image processing, and extracts rich local features by introducing deformable convolution, fuses global context information, and guides the global perception process to pay attention to key detail areas. Meanwhile, the feature expression capability is enhanced through the information interaction mechanism between channels. The design effectively improves the modeling capability of the network for the building structure in a complex scene, retains the detail features while maintaining the global consistency, thereby improving the discriminability and accuracy in the change detection task. Through various convolution operations, edge and conventional features are extracted and fused to output, the sensitivity of the model to contour information is enhanced, attention weights are generated combined with difference feature maps, the feature maps are dynamically adjusted, the model is guided to pay attention to the real change area, the misjudgment caused by background noise is effectively inhibited, the perception capability for the target change area is improved, and the robustness and reliability of the overall detection are enhanced.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing image processing, and in particular relates to a method for detecting building changes in remote sensing images based on interactive attention enhancement. Background Technology

[0002] In recent years, with the significant advancements in Earth observation technology and the growing demand for refined identification results, high-resolution remote sensing imagery, with its rich and detailed surface information, has gradually become an ideal data source for change detection tasks. However, when faced with complex and diverse scenes in imagery, traditional change detection methods, such as those based on algebraic calculations, image transformations, and classification comparisons, often perform poorly.

[0003] In building change detection tasks, detailed local information is often more helpful than global information in identifying changes. However, most current detection methods establish global feature perception by calculating similarity, which often ignores the importance of local information and lacks attention to local features. This leads to a lack of detailed contextual relationships and, consequently, ignores the subtle differences between image pairs of the changed target and its relationship with the environment.

[0004] Meanwhile, when processing dual-temporal remote sensing images, dynamic changes in the scene itself often introduce interference information, such as the construction of temporary facilities and seasonal changes in vegetation. Furthermore, variations in lighting conditions may cause buildings to exhibit different spectral characteristics at different times and from different viewpoints, potentially producing spurious changes. Existing detection methods have limited robustness to such interference and are prone to false detections. Summary of the Invention

[0005] The present invention aims to solve the above problems and provides a method for detecting building changes in remote sensing images based on interactive attention enhancement.

[0006] In a first aspect, the present invention provides a method for detecting building changes in remote sensing images based on interactive attention enhancement, comprising: Step 1: Extract multi-stage features from the input dual-temporal remote sensing images using a weight-sharing twin ResNet18 backbone network. ， ;in n =1, 2, 3, 4; Step 2: After mapping the dual-temporal features, spatial neighborhood encoding of the key features is performed through deformable convolution to obtain features containing rich local details to guide the modeling of global relationships and generate an attention matrix; then, the attention matrix is ​​multiplied element-wise with the value features and added to the original local features to obtain the enhanced feature maps of each branch at multiple stages. Step 3: Perform Sobel convolution on the enhanced feature map to obtain the gradient magnitude, and calculate the difference between the gradient magnitude and the enhanced feature map to obtain the edge features; after convolution and thinning, output the comprehensive feature map; perform dimensionality reduction operation on the comprehensive feature map and its corresponding difference feature map using convolution, then add the two feature maps and process them through an activation function to obtain the weighted feature map; calculate the difference between the weighted feature maps of multiple time phases to obtain the final difference feature map of each stage; Step 4: Perform dimensionality reduction and standardization on the final difference feature maps from each stage; adjust the size of the feature maps using bilinear interpolation; stitch all feature maps together along the channel dimension; perform convolution and ReLU activation operations on the stitched feature maps to generate the final output transformation result map.

[0007] Furthermore, in the remote sensing image building change detection method based on interactive attention enhancement described in this invention, step two, which involves spatial neighborhood encoding of key features using deformable convolution, specifically includes: set up ， Indicates the first n The feature map of the layer, where n =1, 2, 3, 4; H represents the feature map height, W represents the feature map width, and C represents the number of feature map channels.

[0008] Will ， Mapped to query features respectively Q Key features K Sum value characteristics V ; Employing deformable convolution to pair key features K Encoding within the spatial neighborhood yields features that include local correlations. The expression is: ; In the formula, p 0 indicates the position in the input feature map. p n This indicates the position in the output feature map. S Indicates the range of the receptive field. ω ( p n ) and Δ p n These represent the weights and offsets of the deformable convolution, respectively, both generated by an additional convolutional layer: ; In the formula, W offset and b offsetThese represent the weights and biases used to generate the offset convolutional layer, respectively.

[0009] Furthermore, in the remote sensing image building change detection method based on interactive attention enhancement described in this invention, step two, which involves multiplying the attention matrix element-wise with the value features and adding it to the original local features to obtain the enhanced feature maps for each branch at multiple stages, specifically involves: The features of the two branches are concatenated along the channel dimension; the expression is: X cat = Cat ([ Q , K ' ],dim=1); In the formula, Cat This indicates a concatenation operation, where dim=1 represents the channel dimension of the input features; The concatenated features are subjected to two convolution operations to obtain the attention score matrix, expressed as follows: Att= softmax [ Conv ( ReLU ( Conv ( X cat )))]; In the formula, Conv This represents a 1×1 convolution operation; Attention matrix Att AND value characteristics V Perform element-wise multiplication and then combine with the original local features. K ' By adding them together, we finally obtain the enhanced feature maps of each branch. F out : .

[0010] Furthermore, in the remote sensing image building change detection method based on interactive attention enhancement described in this invention, step three involves performing a Sobel convolution on the enhanced feature map to obtain the gradient magnitude, and then calculating the difference between the gradient magnitude and the enhanced feature map to obtain the edge features; specifically: The Sobel operator is introduced to extract edge features, where the gradient values ​​in the horizontal and vertical directions are extracted using two different convolutional layers: ; Next, the feature map F out Perform Sobel convolution operations separately to obtain the gradient maps of the image in the horizontal and vertical directions. F H and F VThen, by calculating the gradient magnitude F G To obtain edge information: ; Step three involves outputting a comprehensive feature map after convolutional thinning; specifically: gradient magnitude F G and input features F out Perform absolute difference calculation: F edge =| F G - F out |; Obtaining edge features F edge Then, a 1×1 convolutional layer is used to... F out Further refinement of feature extraction yields F norm : F norm = Conv norm ( F out ); Finally F edge and F norm Add the values ​​at the corresponding positions to output the comprehensive feature. F I : F I =F norm +F edge .

[0011] Furthermore, in the remote sensing image building change detection method based on interactive attention enhancement described in this invention, step three involves performing dimensionality reduction operations on the comprehensive feature map and its corresponding difference feature map using convolution, then adding the two feature maps and processing them through an activation function to obtain a weighted feature map; the difference between the multi-temporal weighted feature maps is then calculated to obtain the final difference feature map; specifically: Features F I and its corresponding difference feature map F CI Dimensionality reduction is performed using convolutions, then the two feature maps are summed, and a non-linear processing is introduced using the ReLU activation function. ψ raw = ReLU ( Conv ( F I )+ Conv ( F CI )); in, Conv This represents a 1×1 convolution used to integrate feature channel features; Subsequently, the feature maps are fused through a convolutional layer. ψ raw The process yields a single-channel feature map; application... Sigmoid The activation function compresses it to the range [0,1], generating an attention weight map. ψ ; The generated attention weight map ψ With input feature map F I Element-wise multiplication yields the weighted feature maps. F IS : ; Finally, the multi-temporal weighted feature map was analyzed. F IS The difference is calculated to obtain the final difference feature map. D .

[0012] Secondly, the present invention provides a remote sensing image building change detection system based on interactive attention enhancement, comprising: a twin backbone network, a local-global interactive attention module, a change perception attention enhancement module, and a feature fusion decoding output module; The twin backbone network is used to extract multi-stage features from the input dual-temporal remote sensing images through a weight-sharing twin-structure backbone network. ， ;in n =1, 2, 3, 4; The local-global interactive attention module is used to integrate bi-temporal features. and Mapped to query features respectively Q Key features K Sum value characteristics V The key features are spatially encoded using deformable convolution to obtain features containing rich local details, which guide the modeling of global relationships and generate an attention matrix. Subsequently, the attention matrix is ​​multiplied element-wise with the value features and added to the original local features to obtain multi-stage enhanced feature maps for each branch. The change-aware attention enhancement module is used to process the enhanced feature maps at each stage. First, the Comprehensive Feature Extractor (CFE) uses the Sobel operator and absolute difference operation to enhance the illumination robustness of the features at each stage. Then, the Fusion Attention Mechanism (FAM) is used to weight and fuse the initially generated difference features with the comprehensive features output by the Comprehensive Feature Extractor, and the comprehensive feature map is dynamically adjusted by the weights. Finally, the difference between the fused comprehensive feature maps is calculated stage by stage to output the final difference feature maps for each stage. The feature fusion decoding output module is used to perform dimensionality reduction and standardization on the final difference feature maps of each stage; adjust the size of the feature maps by bilinear interpolation; stitch all feature maps together in the channel dimension; and perform convolution and ReLU activation operations on the stitched feature maps to generate the final output change result map.

[0013] Thirdly, the present invention provides a remote sensing image building change detection device based on interactive attention enhancement, comprising a memory and a processor; the memory is used to store a computer program; the processor is used to implement the remote sensing image building change detection method based on interactive attention enhancement as described in any one of the first aspects when the computer program is executed.

[0014] Fourthly, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the remote sensing image building change detection method based on interactive attention enhancement as described in any one of the first aspects.

[0015] The remote sensing image building change detection method based on interactive attention enhancement described in this invention introduces deformable convolution to extract rich local features and fuses global contextual information, guiding the global perception process to focus on key detail regions. Simultaneously, this module further enhances feature representation capabilities through inter-channel information interaction mechanisms. This design effectively improves the network's ability to model building structures in complex scenes, preserving detailed features while maintaining global consistency, thereby improving the discriminativity and accuracy in change detection tasks. Furthermore, by extracting edge and regular features through various convolution operations and fusing the output, the model's sensitivity to contour information is enhanced; attention weights are generated by combining differential feature maps, dynamically adjusting the feature maps to guide the model to focus on truly changed areas, effectively suppressing misjudgments caused by background noise, improving the perception of target change areas, and enhancing the overall robustness and reliability of detection. Combined with multi-level feature fusion decoding, unified modeling and integrated optimization of local details, change perception, and multi-scale information are achieved. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the architecture of the remote sensing image building change detection method based on interactive attention enhancement as described in an embodiment of the present invention; Figure 2 This is a schematic diagram of the architecture of the interactive attention output enhancement feature part according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the architecture of the change-aware attention output difference feature part according to an embodiment of the present invention; Figure 4 This is a schematic diagram comparing the qualitative results of different methods described in the embodiments of the present invention on the S2Looking test set. Detailed Implementation

[0017] The following detailed description of the remote sensing image building change detection method based on interactive attention enhancement, as described in this invention, is provided with reference to the accompanying drawings and embodiments.

[0018] This embodiment discloses a method for detecting building changes in remote sensing images based on interactive attention enhancement, including the following steps: Step 1: Extract multi-stage features from the input dual-temporal remote sensing images using a weight-sharing twin ResNet18 backbone network. ， ;in n =1, 2, 3, 4; Step 2: After mapping the dual-temporal features, spatial neighborhood encoding of the key features is performed through deformable convolution to obtain features containing rich local details to guide the modeling of global relationships and generate an attention matrix; then, the attention matrix is ​​multiplied element-wise with the value features and added to the original local features to obtain the enhanced feature maps of each branch at multiple stages. Step 3: Perform Sobel convolution on the enhanced feature map to obtain the gradient magnitude, and calculate the difference between the gradient magnitude and the enhanced feature map to obtain the edge features; after convolution and thinning, output the comprehensive feature map; perform dimensionality reduction operation on the comprehensive feature map and its corresponding difference feature map using convolution, then add the two feature maps and process them through an activation function to obtain the weighted feature map; calculate the difference between the weighted feature maps of multiple time phases to obtain the final difference feature map of each stage; Step 4: Perform dimensionality reduction and standardization on the final difference feature maps from each stage; adjust the size of the feature maps using bilinear interpolation; stitch all feature maps together along the channel dimension; perform convolution and ReLU activation operations on the stitched feature maps to generate the final output transformation result map.

[0019] In this embodiment of the disclosure, the specific execution process is as follows: First, set I 1 andI 2 represents images of the same area acquired at different times. The architecture of the Remote Sensing Change Detection Network for Buildings with Interactive Attention-Enhanced (IAE-CDNet) is as follows: Figure 1 As shown.

[0020] In IAE-CDNet, dual-temporal remote sensing images The inputs are fed into the ResNet18 backbone network with a weight-sharing twin structure to extract their respective deep features.

[0021] To address the issue of lost local details in most global awareness methods, this embodiment employs a local-global interactive attention approach, the structure of which is as follows: Figure 2 As shown, the dual-temporal features are first extracted through the backbone network. F X and F Y ,set up ， Indicates the first n The feature map of the layer, where n =1, 2, 3, 4. Next, ， Mapped to query features respectively Q (query), key features K (key) and value characteristics V (value); In change monitoring tasks, detailed local information perception can more accurately capture and characterize changes in a scene. However, due to the diversity of building shapes and surrounding environments, models are often limited by the fixed receptive field of traditional convolution operations when capturing these complex local features. Therefore, in this embodiment, deformable convolution is first used to target key features. K Encoding within the spatial neighborhood yields a result containing rich local correlation features. The expression is: ; In the formula, p 0 indicates the position in the input feature map. p n This indicates the position in the output feature map. S Indicates the range of the receptive field. ω ( p n ) and Δ pn These represent the weights and offsets of the deformable convolution, respectively, both generated by an additional convolutional layer: ; In the formula, W offset and b offset These represent the weights and biases used to generate the offset convolutional layer, respectively.

[0022] In this disclosed example, the features of the two branches are first concatenated along the channel dimension. This feature coupling strategy allows one branch to focus on itself while referencing feature information from the other branch, thereby enhancing the network's perceptual capabilities. The operational formula is as follows: X cat = Cat ([ Q , K ' ],dim=1); In the formula, Cat This indicates a concatenation operation, where dim=1 represents the channel dimension of the input features. Subsequently, the concatenated features are subjected to two convolution operations to obtain the attention score matrix, as shown in the following formula: Att = softmax [ Conv ( ReLU ( Conv ( X cat )))]; In the formula, Conv This represents a 1×1 convolution operation. It's worth noting that in... Att In this model, the weight value at each position takes full account of detailed local feature information.

[0023] Finally, the attention matrix containing information about local and global spatial interactions will be used. Att AND value characteristics V Perform element-wise multiplication and then combine with the original local features. K ' By adding them together, we finally obtain the enhanced feature maps of each branch. F out : .

[0024] The edge information of ground features in an image reflects changes in pixel gradients. These edges are less affected by changes in lighting conditions, thus contributing to improved lighting robustness. In this embodiment, a change-aware attention approach is employed, with the structure as follows: Figure 3As shown, the Sobel operator is first introduced to extract edge features. The gradient values ​​in the horizontal and vertical directions are extracted using two different convolutional layers, as detailed below: ; Next, the feature map F out Perform Sobel convolution operations separately to obtain the gradient maps of the image in the horizontal and vertical directions. F H and F V Then, by calculating the gradient magnitude F G To obtain edge information, the formula is as follows: ;

[0025] In this embodiment of the disclosure, in order to further reduce the impact of global brightness fluctuations in the image and enhance the contrast of edge information, the gradient magnitude is... F G and input features F out Perform absolute difference calculation. This yields the edge features. F edge Then, a 1×1 convolutional layer is used to... F out Further refinement of feature extraction yields F norm Finally F edge and F norm Add the values ​​at the corresponding positions to output the comprehensive feature. F I This series of processes is illustrated by the following formula: F edge =| F G - F out |; F norm = Conv norm ( F out ); F I =F norm +F edge。

[0026] In change detection tasks, difference feature maps play a crucial role in both quantitatively and qualitatively identifying areas of change. However, directly performing pixel-level subtraction often introduces noise from irrelevant areas, directly impacting the accuracy of the results. Therefore, it is necessary to establish a correspondence between the difference features and the original features. In this embodiment, a fusion attention mechanism is designed to enable the model to better focus on task-related change features.

[0027] The fusion attention mechanism generates an attention weight map by fusing the two-phase feature maps and the difference feature map between them. φ This involves weighting the original image feature map to enhance its representation of features of interest. Specifically, for the feature map output by CFE... F I and its corresponding difference feature map F CI Dimensionality reduction is performed using convolutions, and then the two feature maps are summed, with a non-linear processing introduced through the ReLU activation function. The formula is shown below: ψ raw = ReLU ( Conv ( F I )+ Conv ( F CI )); in, Conv This represents a 1×1 convolution used to integrate feature channel features.

[0028] Subsequently, the feature maps are fused through a convolutional layer. ψ raw The process yields a single-channel feature map; application... Sigmoid The activation function compresses it to the range [0,1], generating an attention weight map. ψ The generated attention weight map ψ With input feature map F I Element-wise multiplication yields the weighted feature maps. F IS : ; This ensures that the feature value at each location in the input feature map is adjusted according to its corresponding attention weight, thereby enhancing the feature representation of truly changing regions while reducing interference from spurious changing regions. Finally, the multi-temporal weighted feature map is processed. F IS The difference is calculated to obtain the final difference feature map. D .

[0029] To better utilize multi-scale output features and ensure that the transformation results are compatible with spatial detail and semantic information, this embodiment employs a multi-level feature fusion strategy, such as... Figure 1 As shown, let F i =( f 1, f 2, f 3, f 4) To represent the output characteristics at different stages, the following process is executed: F i ' = ReLU ( BN ( Conv (F) i ' ))); F i '' = Interp (F) i ' , size =F ' i+1 ); F cat = Concat (F1) '' F2 '' F3 '' F4 '' ); F out = ReLU ( Conv (F) cat )); In the above formula, through convolutional layers, BatchNorm Layers and ReLU activation functions are used to reduce the dimensionality and standardize the feature maps at different levels. Next, to ensure spatial consistency of the feature maps before concatenation, bilinear interpolation is used to adjust the size of the feature maps at each stage. Then, all feature maps are concatenated along the channel dimension to form a comprehensive feature map. Finally, convolution and ReLU activation operations are performed on the concatenated feature map to generate the final output transformation map.

[0030] This multi-level feature fusion strategy makes full use of the complementarity of features at different stages, which can significantly improve the model's ability to represent the results of complex scene changes.

[0031] In this embodiment, to verify the effectiveness of the method described herein, several advanced techniques in the field of change detection in recent years were selected and compared on the S2Looking dataset. This dataset contains a large number of lighting variations, ground feature differences, and small target buildings. These complex factors pose a significant challenge to the identification of building edges and detailed regions. Therefore, this dataset can be used to focus on evaluating the robustness of the proposed method in dealing with lighting variations and complex background interference, as well as its ability to preserve building detail features.

[0032] To ensure fairness in the comparison, three methods were selected: convolutional neural networks, attention mechanisms, and Transformers. These methods cover the main research directions in current change detection technology, ensuring the comprehensiveness and representativeness of the experimental results. Experimental results demonstrate that the IAE-CDNet network proposed in this invention outperforms many advanced methods in recent years, exhibiting strong detection performance.

[0033] Table 1 shows the quantitative comparison results of IAE-CDNet and the methods mentioned above on the S2Looking dataset. The best results are shown in bold, and the second-best results are shown in underlined. IAE-CDNet's main evaluation metrics, F1 score and IoU, are 75.263% and 60.338%, respectively, representing improvements of 2.772% and 2.218% compared to the second-best method, SUNNet. It is worth noting that this method employs a module that tends to suppress false positives, resulting in a slightly lower recall rate while improving the reliability of the results. However, it achieves the best results in both the F1 score and IoU, indicating that IAE-CDNet has a significant advantage in accurately locating changing regions and avoiding false positives.

[0034] Table 1

[0035] like Figure 4 As shown, a visual comparison of the test results of IAE-CDNet and the methods described above on randomly sampled data on the S2Looking dataset is presented. Figure 4 Rows (a), (c), and (e) illustrate scenes with significant building changes. In (a) and (c), FC-EF suffers from numerous missed detections due to its simple structure, while other methods generally exhibit varying degrees of "adhesion" problems. Because building targets and their surrounding impermeable areas share similar color characteristics, other methods mistakenly identify non-building areas such as roads as building changes. Furthermore, these methods have limited ability to perceive local details and cannot effectively eliminate scene interference, resulting in blurred building boundaries.

[0036] exist Figure 4Rows (b), (d), and (f) illustrate the complexity and significant lighting variations in the ground scenes within the S2Looking dataset. In (b), there are structurally complex buildings with high similarity to the scene. Although both SUNNet and the proposed IAE-CDNet achieve low false positives and false negatives, IAE-CDNet's detection results are more complete. In (d) and (f), the ground scenes are more complex and dense, with significant lighting variations. For (d), other models make relatively conservative predictions, resulting in more false negatives; while for (f), the visual changes in buildings caused by the lighting differences lead to numerous false positives from other models.

[0037] This second embodiment discloses a remote sensing image building change detection system based on interactive attention enhancement, including: a twin backbone network, a local-global interactive attention module, a change perception attention enhancement module, and a feature fusion decoding output module; The twin backbone network is used to extract multi-stage features from the input dual-temporal remote sensing images through a weight-sharing twin-structure backbone network. ， ;in n =1, 2, 3, 4; The local-global interactive attention module is used to integrate bi-temporal features. and Mapped to query features respectively Q Key features K Value characteristics V The key features are spatially encoded using deformable convolution to obtain features containing rich local details, which guide the modeling of global relationships and generate an attention matrix. Subsequently, the attention matrix is ​​multiplied element-wise with the value features and added to the original local features to obtain multi-stage enhanced feature maps for each branch. Existing change detection methods, while capable of some degree of global modeling when processing high-resolution remote sensing imagery, often overlook rich detail features, especially in complex scenes. This can easily lead to the loss of local building structure information, thus affecting detection accuracy. In this embodiment, deformable convolution is first used to acquire rich local feature information, introducing convolutional fusion to guide perceptual global modeling through these rich local features. This strengthens the attention learning process and supports information interaction between channels. LGAM enables the network to better capture architectural details in complex scenes and provides more discriminative features for subsequent processing.

[0038] The change-aware attention enhancement module is used to process the enhanced feature maps at each stage. First, the Sobel operator and absolute difference operation are used in the CFE to enhance the illumination robustness of the features at each stage. Then, the FAM performs weighted fusion of the initially generated difference features and the comprehensive features output by the CFE, and dynamically adjusts the comprehensive feature map through weights. Finally, the difference is calculated stage by stage on the fused comprehensive feature map to output the final difference feature map at each stage. In dual-temporal remote sensing imagery, changes in lighting conditions and non-structural changes in the scene (such as temporary facilities and vegetation changes) can easily introduce spurious changes, interfering with the identification of truly changed targets such as buildings. Existing methods have limited robustness to such interference and are prone to false detections. Therefore, this embodiment employs a change-aware attention enhancement module. The integrated feature extractor extracts edge and regular features through different convolutional operations, then fuses the output to further optimize the LGAM results. The fusion attention mechanism combines the difference feature map to generate attention weights, achieving dynamic adjustment of the original features and enhancing the model's ability to perceive changes of interest.

[0039] The feature fusion decoding output module is used to perform dimensionality reduction and standardization on the final difference feature maps of each stage; adjust the size of the feature maps by bilinear interpolation; stitch all feature maps together in the channel dimension; and perform convolution and ReLU activation operations on the stitched feature maps to generate the final output change result map.

[0040] The remote sensing image building change detection system based on interactive attention enhancement described in this embodiment can effectively retain and utilize local detailed features by fully modeling global context information, thereby enhancing the model's ability to perceive and recognize changes in targets of interest and achieving more refined building change recognition.

[0041] The specific operation steps of the remote sensing image building change detection system based on interactive attention enhancement described in this embodiment are the same as those of the remote sensing image building change detection method based on interactive attention enhancement described in Embodiment 1 above, and will not be repeated here.

[0042] This embodiment three discloses a remote sensing image building change detection device based on interactive attention enhancement, including a memory and a processor; the memory is used to store a computer program; the processor is used to implement the remote sensing image building change detection method based on interactive attention enhancement described in embodiment one when the computer program is executed. The specific detection method steps are the same as those in the aforementioned embodiment one, and will not be repeated here.

[0043] This embodiment four discloses a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the remote sensing image building change detection method based on interactive attention enhancement described in embodiment one. The specific detection method steps are the same as those in embodiment one, and will not be repeated here.

[0044] The computer described in this application embodiment can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. The computer-readable storage medium can be any usable medium that a computer can read, or a data storage device such as a server or data center that integrates one or more usable media. The usable medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., digital versatile optical disc (DVD)), or a semiconductor medium (e.g., solid-state drive (SSD)). The software formed by the computer's stored code can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other storage media that are mature in the art.

[0045] In the various embodiments of this application, the functional modules can be integrated into one processing unit or module, or each module can exist physically separately, or two or more modules can be integrated into one unit or module. In the above embodiments, they can be implemented entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, they can be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated.

[0046] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for detecting building changes in remote sensing images based on interactive attention enhancement, characterized in that... include: Step 1: Extract multi-stage features from the input dual-temporal remote sensing images using a weight-sharing twin ResNet18 backbone network. ， ;in n =1, 2, 3, 4; Step 2: After mapping the dual-temporal features, spatial neighborhood encoding of the key features is performed through deformable convolution to obtain features containing rich local details to guide the modeling of global relationships and generate an attention matrix; then, the attention matrix is ​​multiplied element-wise with the value features and added to the original local features to obtain the enhanced feature maps of each branch at multiple stages. Step 3: Perform Sobel convolution on the enhanced feature map to obtain the gradient magnitude, and calculate the difference between the gradient magnitude and the enhanced feature map to obtain the edge features; After convolutional thinning, a comprehensive feature map is output. The comprehensive feature map and its corresponding difference feature map are then subjected to dimensionality reduction operations using convolution. The two feature maps are then added together and processed by an activation function to obtain a weighted feature map. The difference between the weighted feature maps of multiple time phases is calculated to obtain the final difference feature map for each stage. Step 4: Perform dimensionality reduction and standardization on the final difference feature maps from each stage; adjust the size of the feature maps using bilinear interpolation; stitch all feature maps together along the channel dimension; perform convolution and ReLU activation operations on the stitched feature maps to generate the final output transformation result map.

2. The method for detecting building changes in remote sensing images based on interactive attention enhancement according to claim 1, characterized in that, Step two involves spatial neighborhood encoding of the key features using deformable convolution, specifically as follows: set up ， Indicates the first n The feature map of the layer, where n =1, 2, 3, 4; H Indicates feature map height, W Indicates feature map width, C Indicates the number of channels in the feature map; Will ， Mapped to query features respectively Q Key features K Sum value characteristics V ; Employing deformable convolution to pair key features K Encoding within the spatial neighborhood yields features that include local correlations. The expression is: ； In the formula, p 0 indicates the position in the input feature map. p n This indicates the position in the output feature map. S Indicates the range of the receptive field. ω ( p n ) and Δ p n These represent the weights and offsets of the deformable convolution, respectively, both generated by an additional convolutional layer: ； In the formula, W offset and b offset These represent the weights and biases used to generate the offset convolutional layer, respectively.

3. The method for detecting building changes in remote sensing images based on interactive attention enhancement according to claim 2, characterized in that, Step two involves element-wise multiplying the attention matrix with the value features and adding it to the original local features to obtain the enhanced feature maps for each branch at multiple stages. Specifically: The features of the two branches are concatenated along the channel dimension; the expression is: X cat = Cat ([ Q , K ' ],dim=1); In the formula, Cat This indicates a concatenation operation, where dim=1 represents the channel dimension of the input features; The concatenated features are subjected to two convolution operations to obtain the attention score matrix, expressed as follows: That= softmax [ Conv ( ReLU ( Conv ( X cat )))]; In the formula, Conv This represents a 1×1 convolution operation; Attention matrix Att AND value characteristics V Perform element-wise multiplication and then combine with the original local features. K ' By adding them together, we finally obtain the enhanced feature maps of each branch. F out : 。 4. The method for detecting building changes in remote sensing images based on interactive attention enhancement according to claim 3, characterized in that, Step three involves performing a Sobel convolution on the enhanced feature map to obtain the gradient magnitude, and then calculating the difference between the gradient magnitude and the enhanced feature map to obtain the edge features; specifically: The Sobel operator is introduced to extract edge features, where the gradient values ​​in the horizontal and vertical directions are extracted using two different convolutional layers: ； Next, the feature map F out Perform Sobel convolution operations separately to obtain the gradient maps of the image in the horizontal and vertical directions. F H and F V Then, by calculating the gradient magnitude F G To obtain edge information: ； Step three involves outputting a comprehensive feature map after convolutional thinning; specifically: gradient magnitude F G and input features F out Perform absolute difference calculation: F edge =| F G - F out |； Obtaining edge features F edge Then, a 1×1 convolutional layer is used to... F out Further refinement of feature extraction yields F norm : F norm = Conv norm ( F out )； Finally F edge and F norm Add the values ​​at the corresponding positions to output the comprehensive feature. F I : F I =F norm +F edge 。 5. The method for detecting building changes in remote sensing images based on interactive attention enhancement according to claim 4, characterized in that, Step three involves performing dimensionality reduction operations on the comprehensive feature map and its corresponding difference feature map using convolution, then adding the two feature maps and processing them through an activation function to obtain a weighted feature map; finally, the difference between the multi-temporal weighted feature maps is calculated to obtain the final difference feature map; specifically: Features F I and its corresponding difference feature map F CI Dimensionality reduction is performed using convolutions, then the two feature maps are summed, and a non-linear processing is introduced using the ReLU activation function. ψ raw = ReLU ( Conv ( F I )+ Conv ( F CI ))； in, Conv This represents a convolution of size 1×1; Subsequently, the feature maps are fused through a convolutional layer. ψ raw The process yields a single-channel feature map; application... Sigmoid The activation function compresses it to the range [0,1], generating an attention weight map. ψ ; The generated attention weight map ψ With input feature map F I Element-wise multiplication yields the weighted feature maps. F IS : ； Finally, the multi-temporal weighted feature map was analyzed. F IS The difference is calculated to obtain the final difference feature map. D .

6. A remote sensing image building change detection system based on interactive attention enhancement, characterized in that... include: The system consists of a twin backbone network, a local-global interactive attention module, a change-aware attention enhancement module, and a feature fusion decoding output module. The twin backbone network is used to extract multi-stage features from the input dual-temporal remote sensing images through a weight-sharing twin-structure backbone network. ， ;in n =1, 2, 3, 4; The local-global interactive attention module is used to integrate bi-temporal features. and Mapped to query features respectively Q Key features K Sum value characteristics V The key features are spatially encoded using deformable convolution to obtain features containing rich local details, which guide the modeling of global relationships and generate an attention matrix. Subsequently, the attention matrix is ​​multiplied element-wise with the value features and added to the original local features to obtain multi-stage enhanced feature maps for each branch. The change-aware attention enhancement module is used to process the enhanced feature maps at each stage. First, the feature extractor uses the Sobel operator and absolute difference operation to enhance the illumination robustness of the features at each stage. Subsequently, the initially generated difference features are weighted and fused with the comprehensive features output by the comprehensive feature extractor through a fusion attention mechanism, and the comprehensive feature map is dynamically adjusted through weights. Finally, the difference between the fused integrated feature maps is calculated stage by stage, and the final difference feature maps of each stage are output. The feature fusion decoding output module is used to perform dimensionality reduction and standardization on the final difference feature maps of each stage; adjust the size of the feature maps by bilinear interpolation; stitch all feature maps together in the channel dimension; and perform convolution and ReLU activation operations on the stitched feature maps to generate the final output change result map.

7. A device for detecting building changes in remote sensing images based on interactive attention enhancement, characterized in that: It includes a memory and a processor; the memory is used to store a computer program; the processor is used to implement the remote sensing image building change detection method based on interactive attention enhancement as described in any one of claims 1-5 when the computer program is executed.

8. A computer-readable storage medium, characterized in that: The storage medium stores a computer program, which, when executed by a processor, implements the remote sensing image building change detection method based on interactive attention enhancement as described in any one of claims 1-5.

Images